We have developed a two-level SVM system CanSavPre to predict cancer-related single amino acid variation. Not only protein sequences but also structures are used for descriptors extracted for model training. Our experiment showed much better improvement in the two-level prediction system, and it means more adequate information is necessary for identifying cancer-related SAVs from the divergent sequence of promiscuous protein function in an extensive network of cells. Even though without structure resolved for many sequences, the precise structure information can still be extracted with the help of the homologous search on the PDB database, like homology modeling method. To take into account the properties of the conformation and environment surrounding SAVs, the performance of the result in this work significantly enhanced obviously. Furthermore, the algorithm picked up the optimized the best combination feature vectors using for each kind of variation for specific amino acid type. Therefore, the difference is distanced feasibly.